NLPGuard: A Framework for Mitigating the Use of Protected Attributes by NLP Classifiers

TL;DR

NLPGuard is a framework designed to reduce the reliance of NLP classifiers on protected attributes, effectively decreasing bias and dependence on sensitive information without sacrificing model accuracy.

Contribution

It introduces a method to modify training data to mitigate reliance on protected attributes in NLP classifiers, addressing a gap in bias mitigation techniques.

Findings

Dependence on protected attributes can be as high as 23% in NLP classifiers.

NLPGuard reduces reliance on protected attributes by up to 79%.

It slightly improves classifier accuracy while reducing bias.

Abstract

AI regulations are expected to prohibit machine learning models from using sensitive attributes during training. However, the latest Natural Language Processing (NLP) classifiers, which rely on deep learning, operate as black-box systems, complicating the detection and remediation of such misuse. Traditional bias mitigation methods in NLP aim for comparable performance across different groups based on attributes like gender or race but fail to address the underlying issue of reliance on protected attributes. To partly fix that, we introduce NLPGuard, a framework for mitigating the reliance on protected attributes in NLP classifiers. NLPGuard takes an unlabeled dataset, an existing NLP classifier, and its training data as input, producing a modified training dataset that significantly reduces dependence on protected attributes without compromising accuracy. NLPGuard is applied to three classification tasks: identifying toxic language, sentiment analysis, and occupation classification. Our evaluation shows that current NLP classifiers heavily depend on protected attributes, with up to $23\%$ of the most predictive words associated with these attributes. However, NLPGuard effectively reduces this reliance by up to $79\%$, while slightly improving accuracy.